Spatial Distribution Feature Learning for DOA Estimation in Multi-path Environment

Multi-path effect has negative effects on DOA estimation, especially for low-elevation targets for very-high-frequency (VHF) radar. It also causes the problem of model mismatch and degrades the performance of parametric approaches. Therefore, in this study, we present an approach that learns the spatial distribution feature by an unsupervised auto-encoder. The neural network is trained in the training stage, and in the test stages, the spatial distribution feature is learned and projected into the feature set and thus reverse the DOA. Experimental results indicate that the proposed approach outperforms the classic parametric approaches in DOA estimation accuracy and has a higher adaptability in real-world environment.